Wind Turbine Major Components Failure

Predicting Based on SCADA Data Analysis

Li Shaowu

EWEA Analysis of Operating Wind Farms 2016

Spain Bilbao Apr 2016

China Longyuan Power Group

Xu Jia

Liu Ruihua

City University of Hong Kong

Zhang Zijun

Wang Long

Long Huan

Co-authors

>15 GW Wind Energy >11000 Wind Turbines >160Wind Farms

22 Different Turbine Manufacturers 90 Models

Investment, Design, Development, Construction, Management, O&M service.

Owner maintained more than 75% turbines in China

LongYuan Power Wind Energy Overview/Portfolio

Turbine by turbine analysis

Benchmark analysis of wind farms KPI

Benchmark analysis of Provincial company

KPIGroup

Liaoning

Heping … Changtu

… Canada

Prufflin

What we use SCADA data for

KPI PBA

MTBT

…

MTBR

•Automatic classification

•Anomaly detection

•Anomaly data mining

Power Curve Monitoring

•Blade

•Gearbox

•Generator

•Converter

Failure Predicting and Monitoring

Major Component Failures in LongYuan，2015

Distribution of Major Component

Failures The damaged blade

The damaged Gearbox

Based on the classical methods and physical rules

Time sequence of oil temperature analysis

Vibration signals analysis Lubricant pressure analysis

Transmission ratio analysis

Failed

Root Cause analysis

•Gearbox failure

Problem

•Vibration

•Lubricant

Key point•CMS

•SCADA

Data

•Correlation

•Model

Method•Predict

•Monitoring

Monitor

Lubricant

pressure

viscosityTemperature

WT Power parameters

Power output

Oil temperature

Shaft temperature

Pump power

Other design paremeters

density

velocity

Resistance coefficient

Lubricant parameters

The effect of iron scrap

Lubricant pressure

viscosityTemperature

WT Power

Power output

Oil temperature

Shaft temperature

Pump power

Other design paremeters

density

velocity

Resistance coefficient

Lubricant system

Iron scrap

1、The lubricant is used to cool down the gearbox

2、 The change of the lubricant pressure follows the change of the power output.

Compared with the gearbox oil temperature, the lubricant pressure is more resistant to

the impact of the environmental temperature.

3、 If there is the mechanical wear on the gearbox, the iron scrap will fall into the

lubricant oil and the lubricant pressure will change.

Therefore, the lubricant pressure, Pl, is chosen as the predictive objective.

The model for Gearbox – Deep Neutral Network

Gearbox oil temperature, To

Power output, Po

Shaft temperature, Ts.

The schematic diagram of DNN

The model for Gearbox – Deep Neutral Network

The trained model is

The activation function:

The training process of DA is to learn the parameters,

0ˆ ( , , )

l s oP f T T P

tanh( )t t

t t

e et

e e

Deep Neutral Network

EWMA Control Chart1e (1 )t rt tz z

2[1 (1 ) ]( )

(2 )r r

t

e eUCL t Ln

2[1 (1 ) ]( )

(2 )r r

t

e eLCL t Ln

The upper and lower EWMA control limits depend on time

2

1

1 ˆ{ } argmin ( ) , 0,1, ,2 l l

N

n n

i

P P n L

n nW ,b

W ,b

Cases - Gearbox

The EWMA chart of normal and abnormal turbines

581523465407349291233175117591

0.20

0.15

0.10

0.05

0.00

-0.05

-0.10

Sample

EW

MA __

X=0.05

UCL=0.2

LCL=-0.1

613545477409341273205137691

0.20

0.15

0.10

0.05

0.00

-0.05

-0.10

Sample

EW

MA __

X=0.05

UCL=0.2

LCL=-0.1

561505449393337281225169113571

0.6

0.5

0.4

0.3

0.2

0.1

0.0

-0.1

Sample

EW

MA

__X=0.05

UCL=0.2

LCL=-0.1

771694617540463386309232155781

25

20

15

10

5

0

Sample

EW

MA

__X=7.81

UCL=18.04

LCL=-2.41

EWMA Chart of error

Wind Farm TurbinesTime prior to

failures

Liaoning No 49 Two days

Hebei No 64 Three days

Question: Can this method be used for Blade?

CORE I5 CPU and 8GB memory

training time: 3-min,

applying time: 1-min.

The model for Blade – deep autoencoder

Schematic diagram of Deep Autoencoder

Cases - Blade

9108097086075064053042031021

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0.0

Sample

EW

MA __

X=0.35

LCL=0.0747

UCL=0.6253

9108097086075064053042031021

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0.0

Sample

EW

MA

__X=0.35

UCL=0.6253

LCL=0.0747

9108097086075064053042031021

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0.0

Sample

EWM

A

__X=0.326

UCL=0.6274

LCL=0.0246

9108097086075064053042031021

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0.0

Sample

EW

MA

__X=0.326

UCL=0.6274

LCL=0.0246

The EWMA chart of normal and abnormal turbines

Wind farm Turbines Time prior to

failures

Shandong No. 9 7 hours and 20

minutes

Shandong No. 2 6 hours and 10

minutes

Anhui No. 5 8 hours

All validation are blind.

CORE I5 CPU and 8GB memory

training time: 40-min,

applying time: 10-min.

• The DNN model is applicable to identify impending gearbox failure base on SCADA data.

• The DA model is applicable to identify impending blade failure based on SCADA data.

• The proposed method raised alarms early enough for the replacement or repair.

• There were no false alarms for failure monitoring.

• The effectiveness of these methods needs to be further examined by more cases.

• At present, these models have been deployed for monitoring the failure in a wind farm of

Longyuan.

Conclusions

Question?

Li Shaowu

lishaowu@clypg.com.cn

李韶武 Li Shaowu

LongYuan(Beijing) wind power engineering technology Co., LTD

Adr：The 6-9th, North Avenue Fuchengmen Xicheng District, Beijing

Tel：010-63887171

Mail：lishaowu@clygp.com.cn